Python · Apache-2.0
youssofal/MTPLX
The fastest way to run Qwen 3.8 Flash Next, Qwen 3.8 27B and Ternary Bonsai 2 27B on a Mac: 125 tok/s in OpenCode on an M5 Max, and a 27B model on 16 GB Macs. Native MTP speculative decoding on Apple Silicon, exact at any temperature. OpenAI and Anthropic compatible local server.
★ 2.7k
Stars
210 forks · 114 open issues
+2.9/day
Growth · measured
Stars counted between two of our own snapshots roughly a week apart. The repository page gives the exact span.
35.7
Momentum / 100
see the breakdown below
6mo ago
Created
last push 7d ago
Star history
How this score was built
| Component | Weight | Earned | Why |
|---|---|---|---|
| Velocity | 55 | 11.2 | +2.9/day, log-scaled against a 750/day ceiling |
| Acceleration | 15 | 0.0 | last week ran 0.66× the monthly pace |
| Freshness | 15 | 14.6 | last push 7d ago; decays to zero at 180 days |
| Scale | 15 | 9.9 | 2.7k stars, log-scaled against 150k — stops a 3-star repo topping the board |
| Momentum | 100 | 35.7 | sum of the above |
Detail
| Topics | anthropic-compatibleapple-siliconclaude-codeflash-nextinference-enginellm-inferencelocal-ailocal-llmmacosmetalmlxmtpmtplxnative-mtpopenai-compatibleqwenqwen3-8qwen3-8-flash-nextqwen3-nextspeculative-decoding |
| Matched on | LLM & Inference (llm-inference) |
| Short window | +20 stars over 7 days = 2.9/day |
| Long window | +113 stars over 26 days = 4.3/day |
| Lifetime average | 14.27 stars/day over 187 days |
| First seen here | 2026-10-11 |
| Last read | 2d ago |